bugfix: fix the problem of incorrect kv cache data format when enabling prefix cache feature(for deepseek v3 model).

Author: longhui-z

xllm/core/runtime/worker_impl.cpp

@@ -287,8 +287,8 @@ bool WorkerImpl::allocate_kv_cache(
         // Full attention layer: allocate key_cache and value_cache only
 #if defined(USE_NPU)
         aclFormat npu_format_type =
-            context_.get_model_args().model_type().compare(
-                0, strlen("deepseek_v3"), "deepseek_v3") == 0 &&
+            absl::StartsWith(context_.get_model_args().model_type(),
+                             "deepseek_v3") &&
                     FLAGS_enable_prefix_cache
                 ? ACL_FORMAT_FRACTAL_NZ
                 : ACL_FORMAT_ND;

xllm/models/llm/npu/mtp_model_base.h

@@ -138,8 +138,7 @@ class MtpModelImplBase : public torch::nn::Module {
         }
         attn_mask = torch::cat(req_mask_vec, 0);
       }
-    } else if (model_type_.compare(0, strlen("deepseek_v3"), "deepseek_v3") ==
-                   0 &&
+    } else if (absl::StartsWith(model_type_, "deepseek_v3") &&
                FLAGS_enable_prefix_cache &&
                !input_params.batch_forward_type.is_decode()) {
       attn_mask =